A TAXONOMY AND RISK-AWARE CONCEPTUAL FRAMEWORK FOR AGENTIC AI-BASED AUTONOMOUS TASK SELECTION IN SOFTWARE ENGINEERING WORKFLOWS

Abstract

Agentic artificial intelligence is changing software engineering assistance by shifting from immediate response to code generation to systems that understand the context, use tools, observe feedback, and decide on follow-up actions. Most AI programming assistants and software agents today focus on tasks like code completion, debugging, testing, or issue handling at repository level, without considering task selection as a distinct, explainable, and measurable decision process layer. In this paper, we propose a taxonomy and risk-aware approach to the concept of agentic AI for autonomous task selection in software engineering processes. Our framework uses developer input and workflow signals for task classification, computes uncertainty and risk estimates, makes decisions about selecting the next suitable action, and refers uncertain or critical cases to human confirmation/clarification. We contribute to literature through a task taxonomy definition, comparison with prior research and gap analysis, design of next action selection architecture, decision policy proposal, and outline of experiment scenarios. The present study represents a survey/conceptual framework type of contribution and will be developed later in prototype-based evaluations.